Anthropic Quietly Ends Zero Data Retention for Its Top Models

1. Anthropic Quietly Ends Zero Data Retention for Its Top Models

TL;DR: Anthropic has been requiring thirty-day data retention for its most capable models — Claude Fable 5 and Claude Mythos 5 — since June 9th, across every platform including Bedrock and Vertex AI, to support internal safety review. OpenAI responded this week with "Private Safety Processing," a mechanism that scans customer traffic for safety signals without exposing raw content to OpenAI staff, explicitly preserving zero data retention.

Key Points:

• Anthropic's "Covered Models" (its frontier tier) now retain prompts and outputs for thirty days regardless of platform — a reversal of the zero-data-retention guarantee many enterprise customers built their procurement decisions around
• Anthropic's own report acknowledges the change "will be unpopular with customers who have come to expect zero retention, and pose real risks to our business success (especially if competitors do not follow)" — a rare public admission of self-inflicted competitive risk
• Non-Covered-Model Anthropic API traffic is unaffected — standard thirty-day-max deletion continues as before
• OpenAI's Private Safety Processing (announced Aug 20) scans customer interactions — including tool-use and agent activity, not just single prompts — for safety signals while keeping raw content unreadable to OpenAI staff; in pilot with Databricks and Microsoft, general availability planned for September
• OpenAI's mechanism is enterprise and API-tier only, not consumer ChatGPT

Deep Dive: For most of the last two years, "Anthropic doesn't retain your data, OpenAI might" was a settled fact enterprise security teams used to make vendor decisions — especially in regulated industries where a model provider quietly logging prompts is a compliance problem, not just a preference. That fact just inverted, and it inverted quietly: no splashy announcement, just a policy document acknowledging the company expects to lose customers over it.

That's the part worth internalizing. This isn't "don't trust Anthropic" — it's "stop trusting the vendor label and start reading which model tier a zero-retention clause actually covers." OpenAI's response is instructive too: Private Safety Processing is a genuinely different architecture choice — trying to get the safety-monitoring benefit of retained data without actually retaining the readable content — and it's a bet that the market will pay a premium for provable non-retention over Anthropic's "we need the data to do safety review properly" argument. Worth watching whether Anthropic's framing (safety review requires retention) or OpenAI's framing (we can do safety review without it) wins the enterprise conversation over the next two quarters — this is also the same trust surface the Google Antigravity story below is retrofitting audit controls onto, and the same territory Anthropic's own multi-agent-sabotage research (covered here Monday) was probing from the "what happens when nobody's watching the agent" angle.

So What? If you're running Claude's frontier models in a regulated or IP-sensitive environment, check which specific model tier your zero-data-retention contract actually covers — "Anthropic" is no longer a safe shorthand for "doesn't retain," and that distinction needs to show up in vendor-risk documentation this quarter, not whenever the next audit forces the question.

SourcesThe Register — OpenAI Chases Anthropic's Biz Customers With Zero Data Retention Pledge, Axios — OpenAI Previews Zero-Retention Safety System as Anthropic Requires Data Logs, Bloomberg — Anthropic Plans to Change Data Retention Policy for Advanced AI

2. Infrahub Goes GA With Selective Regeneration — Merges Stop Taking Coffee Breaks

TL;DR: OpsMill shipped Infrahub v1.11.0 out of beta on August 19th. The headline feature: merges and repo commits now trigger proportional regeneration work instead of full re-runs, cutting validated regeneration task counts by seventy-three to one hundred percent — plus real operability wins around webhook delivery and failed-merge recovery.

Key Points:

• selective_execution_after_merge (env var INFRAHUB_SELECTIVE_EXECUTION_AFTER_MERGE) is on by default, extending 1.10's artifact-level precision up to Generators and computed attributes
• Breaking change: artifacts tied to a removed target now get deleted on the next full regeneration pass — capture content first if you need it
• Breaking change: the infrahub git-agent CLI command is gone entirely, replaced by the task worker — anyone with cron or CI hooks calling that command breaks silently on upgrade until they read the release notes
• Generated OpenAPI clients need regeneration (component schema names changed); container image bumps to Python 3.14 and Neo4j 2026.05.0, so a separately managed Neo4j needs a matching upgrade
• Sustained overload now returns a proper 429 plus Retry-After header — the SDK and UI handle it automatically, but custom API integrations need to add that handling themselves
• A same-day patch, v1.10.9, fixed a Tasks-page crash on unresolvable related nodes and an attribute-rename/rebase duplication bug

Deep Dive: This is the clearest evidence yet that Infrahub is chasing NetBox and Nautobot on operational maturity rather than just leaning on its graph-database schema flexibility as a selling point. Selective regeneration sounds like a minor optimization until you've actually sat through a large-repo merge waiting for every Generator and Transformation to re-run against the whole dataset — a seventy-three to one hundred percent reduction in that task count is the difference between "merge and get coffee" and "merge and go to lunch."

The git-agent removal is the real trap in this release, and it's exactly the kind of breaking change that doesn't show up until three weeks after the upgrade when someone's cron job silently stops working. If you dismissed Infrahub a year ago as "too graph-database-y for production," this release — GA selective regeneration, proper backpressure signaling, real recovery from failed merges — is worth a second look.

So What? If you're running Infrahub with any non-trivial set of Generators or Transformations, upgrade to v1.11.0, but audit every cron job, CI hook, and script that calls infrahub git-agent first — that command is gone, not deprecated, and the failure mode is silent.

SourcesGitHub — Infrahub v1.11.0 Release Notes, GitHub — Infrahub v1.10.9 Release Notes

3. Nineteen Companies Launch an Open Standard for AI-Cluster Optics — A Direct Hedge Against NVLink-Style Lock-In

TL;DR: A nineteen-company coalition led by photonics startup Lightmatter — including Dell, Qualcomm, Foxconn Interconnect, Celestica, Flex, and Keysight — launched Open Silicon Photonics for AI Systems within the Open Compute Project on August 17th, publishing an architecture-vision white paper for vendor-interoperable co-packaged optics across AI clusters from seventy-two nodes up to over a thousand.

Key Points:

• Founding members span the full supply chain: optics (Lightmatter), ODMs (Celestica, Foxconn, Quanta), and test/measurement (Keysight) — a real cross-section, not just marketing partners bolted onto a press release
• Target scale: interoperability across AI clusters from seventy-two nodes to one thousand twenty-four-plus nodes
• The opening artifact is a published architecture-vision white paper, not just an announcement — a genuine attempt to define the interop layer before any single vendor locks it down
• Co-packaged optics itself isn't new — NVIDIA's had Quantum-X and Spectrum-X Photonics on its roadmap for over a year — the news is an open, multi-vendor standard effort specifically aimed at keeping optical I/O from becoming the next NVLink-style proprietary chokepoint

Deep Dive: This is exactly the kind of story this show keeps saying enterprise shops underrate: open standards work happening quietly at the infrastructure layer while everyone's attention is on model releases. NVIDIA's proprietary optical roadmap already works and already ships — that's the competitive pressure forcing everyone else in the supply chain to coordinate on an open alternative before NVLink's interconnect advantage becomes permanent the way it arguably already has for GPU-to-GPU fabric.

The honest caveat: OCP working groups have a mixed track record of turning "architecture vision" white papers into shipping, interop-tested silicon on schedule. This is a multi-year bet, not a 2027 procurement decision — but it's the right bet to be watching, because the alternative is every non-NVIDIA AI cluster buyer eventually facing the same lock-in dynamic that's already reshaping how hyperscalers plan multi-year silicon commitments.

So What? Add "Open Silicon Photonics for AI Systems" to your OCP-tracking list and revisit it at the next OCP Global Summit — if you're specifying AI fabric interconnect on any multi-year horizon, an open CPO standard existing (or not) changes your vendor-lock-in exposure materially.

SourcesDesign & Reuse — Industry Coalition Announces Open Silicon Photonics Initiative Within Open Compute Project

No new protocol, microsegmentation, or SONiC-ecosystem movement cleared the bar this cycle — checked IETF bess/idr/spring/nvo3 draft activity, NANOG/RIPE archives, and SONiC community channels directly against the past week's coverage; nothing genuinely new since Monday's EVPN-over-SRv6 and RIFT Auto-EVPN drafts. This issue's architecture news is the Open Silicon Photonics coalition above — a fair trade for a quiet protocol week.

(Infrahub's v1.11.0 GA is this issue's Top 3 lead for automation — see above.)

Containerlab Is Quietly Hardening Podman Into a First-Class Runtime

TL;DR: Over the past three days, containerlab merged a cluster of Podman-runtime fixes — retrying management-network cleanup on destroy, preserving image command defaults, supporting none network mode — alongside unrelated but notable additions for generic volume mounts and configurable node hostnames.

Key Points:

• PRs #3292, #3278, and #3279 fix Podman-specific runtime bugs; #3335 adds generic volumes support; #3281 adds configurable hostnames — all merged to main on August 19th and 20th
• None of this has landed in a tagged release yet — the latest tag is still v0.78.2 from August 13th
• Two open dependency bumps for go.podman.io/image and go.podman.io/common suggest an active, ongoing push to keep the Podman backend current, not a one-off fix

So What? If you're on a rootless or RHEL-family host, or running containerlab inside CI where Docker-in-Docker is politically or operationally off the table, this is worth watching for the next tagged release rather than acting on today — Podman is close to being a genuinely supported first-class path, not just "technically works."

SourcesGitHub — containerlab Pull Requests

The Git Trick Every Config-as-Code Team Should Already Know

TL;DR: ipSpace.net's latest covers Git forensics techniques — git log -S/-G pickaxe search and reflog-based recovery — applied to infrastructure-as-code repositories, where git blame alone rarely explains intent.

Key Points:

• git log -S<string> finds the exact commit where a string was introduced or removed — turns "who broke the ACL and why" from Slack archaeology into a thirty-second command
• git log -G<regex> does the same with regex matching for less exact searches
• Reflog-based recovery covers the "I force-pushed over something important" scenario every automation team eventually hits

So What? If your NetBox- or Nautobot-backed config generation runs through Git and you've never used pickaxe search, add it to your runbook this week — it's the highest-value five-minute skill upgrade in this issue.

SourcesipSpace.net

(Anthropic's data-retention reversal is this issue's Top 3 lead for AI/ML — see above.)

Google Bolts Enterprise Governance Onto Its Agentic Coding Tool

TL;DR: Google folded Antigravity, its agentic coding assistant, into Gemini Enterprise subscriptions this week, finally adding the admin controls enterprises had been demanding since May's launch: audit logging, file-access restriction to specific directories, terminal-command approval and sandboxing, browser-access allowlisting, and project-level budget caps.

Key Points:

• Antigravity now ships as an extension for VS Code, Visual Studio, JetBrains, and Zed instead of requiring its own standalone IDE
• Per-user and per-team budget controls are explicitly "coming later this year" — only project-level caps are live today
• Lands amid what The Register frames as a broader "AI shakeup" at Google — chief scientist Jeff Dean's departure and DeepMind CEO Demis Hassabis moving to chairman while Koray Kavukcuoglu takes an SVP role, read as DeepMind being pulled closer under direct Google Cloud control

So What? The evaluation criterion for any agentic coding tool — Antigravity, Copilot Workspace, Cursor's enterprise tier — isn't capability anymore, it's whether the audit trail and blast-radius controls actually match what your change-management process requires before an autonomous agent gets shell access to anything production-adjacent. If you're piloting one of these, ask for the directory-scoping and command-approval feature list before the demo, not after.

SourcesThe Register — Google Tethers Antigravity to Enterprise Controls Amid AI Shakeup, Google Cloud Blog — Expanding Google Antigravity for Enterprise Customers

Dell'Oro Nearly Doubles Its Own Capex Forecast to Three Trillion Dollars by 2030

TL;DR: Dell'Oro Group's latest forecast puts global datacenter capex above three trillion dollars by 2030 — nearly double the firm's own January 2026 projection — with high-end AI accelerators the largest spend category and a newly tracked "AI-specialized cloud" segment (model builders and neoclouds, distinct from traditional hyperscalers) growing at almost sixty percent compounded annually.

Key Points:

• Published August 18th; the top four US hyperscalers account for roughly half of global capex
• AI accelerators are projected to account for about a third of the three trillion dollars
• The number itself matters less than the revision velocity: a forecasting house nearly doubling a four-year-out projection in seven months signals the analyst community is still chasing the curve upward, not converging on a stable estimate
• Separately, Data Center Knowledge examined large-load flexibility — dynamically throttling datacenter draw during grid peak-stress periods instead of building fixed capacity for worst-case peak — as a genuinely distinct angle from this month's FERC interconnection-deadline story, framed against the structural mismatch between five-to-seven-year interconnection queues and one-to-three-year hyperscale build cycles

So What? Treat the three-trillion-dollar figure as a moving target, not a plan input — for anyone doing multi-year power or interconnect planning, the revision velocity is the actionable signal: keep those horizons short and reversible rather than betting on any single vendor's 2030 roadmap holding still.

SourcesDell'Oro Group — AI Buildout Maintains Momentum as Data Center Capex Surpasses $3 Trillion by 2030, Data Center Knowledge — AI Infrastructure Pushes Data Center Capex Forecast Above $3 Trillion, Data Center Knowledge — Can Large-Load Flexibility Ease Data Center Energy Concerns?

THE FUN ONE: Your Neighbor's Garage Might Be a Datacenter Now

TL;DR: Startups including Nanocenter.ai, SPAN's XFRA, and Sunrun are piloting wall- and garage-mounted GPU appliances in private homes, orchestrated by an upstream datacenter, as a way to source AI inference capacity without building new hyperscale campuses — a direct, if odd, response to the community pushback that's already killed or delayed hyperscaler projects this year.

Key Points:

• Nanocenter.ai launched in June 2026: RTX 6000-class GPUs installed in homes, homeowners can earn up to two thousand dollars a month, five-year equipment financing, a ten percent platform fee
• Explicitly positioned for inference and batch workloads only — residential network links can't support the tightly-coupled multi-node interconnect that training requires
• Named unsolved problems, in the companies' own framing: hardware attestation and secure boot across thousands of untrusted residential nodes, signed update distribution at fleet scale, and wildly inconsistent home power and network quality

So What? This stays a niche inference-offload play for years, not a real datacenter alternative — the moment any workload needs meaningful inter-node bandwidth, the residential last mile kills it. But it's a legitimate distributed-systems problem (remote attestation at fleet scale, WAN-tolerant orchestration) wearing a homeowner side-hustle costume, and worth watching purely for how the attestation problem gets solved, since that solution generalizes to any untrusted-edge compute scenario.

SourcesData Center Knowledge — Home-Based GPU Networks: Viable Supplements to AI Data Centers?

A Star Fast Enough to Measure a Black Hole's Spin

TL;DR: Using ESO's Very Large Telescope Interferometer, astronomers discovered S301 — a star completing a full orbit around the Milky Way's central black hole, Sagittarius A*, in just eight point seven years, the shortest known period yet, swinging within about one hundred forty Schwarzschild radii at over eight percent of the speed of light.

The Science: Peer-reviewed, published in Nature on August 19th (MPE-led collaboration). Detected via GRAVITY interferometry — four 8.2-meter telescopes phased together — in spring 2023, confirmed against archival data back to 2017. Sagittarius A*'s mass is already well constrained, but its spin has never been directly measured; S301's orbit is tight enough that frame-dragging effects from the black hole's spin should become measurable within roughly a decade using the same instrument.

Why It's Interesting: This would be the first spin measurement of a quiescent — non-actively-feeding — supermassive black hole, opening a strong-field test of general relativity that no prior stellar orbit has reached. It's also a clean demonstration of what large-baseline optical interferometry can resolve at galactic-center distances, twenty-six thousand light-years out.

SourcesNature — Discovery of a Star Sensitive to the Spin of Sagittarius A*, ESO Press Release, MPE News

The Boring Manufacturing Fix That Might Actually Matter for Quantum Hardware

TL;DR: Cornell researchers found that swapping the sputtering gas from argon to krypton lets tantalum films for superconducting qubits form the correct crystal phase at two hundred degrees Celsius instead of over four hundred — low enough to fall within standard semiconductor foundry tooling, without destructive intermixing with the silicon substrate.

The Science: Peer-reviewed, Nature Materials, published August 17th (Valla Fatemi's lab, first author Maciej Olszewski). Krypton's higher atomic mass transfers more momentum during deposition, stabilizing the desired crystal phase at a much lower substrate temperature; the resulting qubits reportedly retain top-tier coherence performance.

Why It's Interesting: Tantalum is currently one of the best-performing materials for superconducting qubits, but its high deposition temperature has kept it largely confined to bespoke lab fabrication rather than standard commercial nanofab lines. This isn't a new qubit architecture — it's the unglamorous process fix that determines whether quantum hardware can scale in volume instead of staying a one-off lab curiosity, and it's exactly the kind of story that doesn't make headlines but actually moves the timeline.

SourcesNature Materials — Krypton-Sputtered Tantalum Films for Scalable High-Performance Quantum Devices, Cornell Chronicle — New Ingredient in Quantum Computing: Krypton Gas

• Security architecture: no significant updates this cycle — resuming checks after Wednesday and Thursday's outage (see below), so this isn't a continuation of last week's streak, just a fresh clean check against CISA, NIST CSRC, Cloud Security Alliance, Risky Business, Krebs, and vendor architecture blogs (Cloudflare, Palo Alto Unit 42, Microsoft Security Research Center).
• Housekeeping: this show didn't run Wednesday or Thursday — an expired login session on our end killed the pipeline before it could research anything, and nobody caught it until this morning's fix. Genuinely embarrassing timing given today's lead story is about AI vendors and trust. We checked back three to four days across every domain this issue to cover the gap rather than just picking up where Tuesday left off.
• A cluster of AI-agent research trended on Hugging Face's papers page this morning — training-environment generation, multi-turn policy compliance for agents, and long-context inference-efficiency work all showed up together. Common thread: making agents safe and cheap enough to run unsupervised at scale. Worth a glance, not yet a full story — treat any specific benchmark numbers as unverified until we can trace them to a primary paper.

SourcesHugging Face Papers

• Whether Anthropic's or OpenAI's framing on data retention versus safety review wins the enterprise procurement conversation over the next quarter
• Fallout from Infrahub's git-agent CLI removal — this is the kind of breaking change that surfaces weeks after upgrade, not on day one
• Whether containerlab's Podman fixes land in a tagged release, and how close v0.79 gets to first-class Podman support
• Google's per-user/per-team Antigravity budget controls, still slated for "later this year"
• Next week: whether this show's own infrastructure stays up two days in a row

• Domains researched: 5 parallel research agents (network architecture + datacenter, network automation, AI/ML, security, science), each widening their search window to 3-4 days to cover Wednesday and Thursday's outage
• Research passes: ~28 web searches and direct GitHub/PyPI/journal checks across domains (network architecture ran 35 direct tool uses, AI/ML ran 16, automation ran 18)
• Major items published: 11 (3 Top 3 + 8 domain items), plus 3 quick takes
• Quality score: 4/5 — strong cross-domain lead (the Anthropic/OpenAI data-retention flip connects directly to this week's Antigravity-governance and multi-agent-trust threads) and automation held its usual weight with a genuine Top-3 win in Infrahub's GA; docked slightly for two sourcing items resting on lower-confidence links (the Hugging Face papers cluster and the Data Center Knowledge flexibility piece) and for the two-day coverage gap itself